KG2Code: Transforming Knowledge Graphs into Code for LLM Question Answering
A new approach called KG2Code converts knowledge graphs into executable code, preserving structural semantics while aligning with code-aware pretraining of large language models. The method addresses limitations in existing KGQA techniques, including structural information loss, unfaithful reasoning, and limited flexibility. KG2Code-QA, a framework built on KG2Code, formulates knowledge graph question answering as a code generation task. The research is published on arXiv under identifier 2607.22652.
Key facts
- KG2Code transforms knowledge graphs into code-based representation
- Preserves structural semantics of knowledge graphs
- Aligns with code-aware pretraining of modern LLMs
- KG2Code-QA formulates KGQA as code generation task
- Addresses limitations of RAG-based, agent-based, and SPARQL-based methods
- Published on arXiv with ID 2607.22652
- Aims to enhance LLM performance on knowledge-intensive tasks
- Proposed to overcome structural information loss, unfaithful reasoning, and limited flexibility
Entities
Institutions
- arXiv